home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!IAF.UQUEBEC.CA!IMR2027
- Organization: Institut Armand Frappier
- Message-ID: <STAT-L%92073009024806@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Thu, 30 Jul 1992 08:52:21 EDT
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- Comments: Adresse alternative de l'envoyeur: <IMR2027@UQUEBEC>
- From: Eduardo Franco <IMR2027@IAF.UQUEBEC.CA>
- Subject: logistic regression
- X-To: stat-l@vm1.mcgill.ca
- Lines: 72
-
- Quoting from lawhiteside%UALR.EDU@VM1.MCGILL.CA...
-
- LW> I have a variable that I think is a very good predictor of my good/bad
- LW> outcome measure. However, it is so good (and I have so few cases)
- LW> that all of the good outcome cases measure in the "good" category
- LW> of the predictor variable ... that is the situation is like
- LW> this (where the numbers are the N in each cell) :
- LW>
- LW>
- LW> Outcome Variable
- LW> Predictor var bad good
- LW> low 143 28
- LW> high 25 0
- LW>
- LW>
- LW> What I really want is to compute odds ratio - but I don't know how to handl
- LW> the 0 cell .....
-
- Add 0.5 to each cell frequency in the above crude table (if I'm not
- mistaken, this is known as Woolf's technique) if you must derive the OR
- as measure of association.
-
- LW> On a related issue, I have another variable in the same model that is
- LW> "the number of people living in the home" ... I want to be able to say
- LW> something like: "Living with 4 or more people puts a child at risk of xxx
- LW> Is there a standard way to determine the cut point? I know I could leave i
- LW> continuous but in the first place I don't think it is linear and in the sec
- LW> I want to know when the risk starts not just that more is worse....
- LW>
- LW> Thanks in advance,
- LW> Leanne Whiteside
- LW> LAWHITESIDE@UALR.EDU
-
- Build your model using dummy covariates to represent contrasts. This method
- comes in handy when calculating ORs. For instance, assume that your variable
- for no. of people is named NUMP. The following set of dummy regressors would
- be used in place of NUMP in your LR model: (assuming hypothetical values)
-
- Value for Coding for dummy variables
- NUMP NUMP2 NUMP3 NUMP4 NUMP5 Level
- 0-1 0 0 0 0 1
- 2-3 1 0 0 0 2
- 4-5 0 1 0 0 3
- 6-7 0 0 1 0 4
- >=8 0 0 0 1 5
-
- Observe that for j levels in your original variable you need to
- specify j-1 dummy regressors. The exponentiated regression coefficient
- for NUMPi will denote the OR for level i contrasted with level 1, taken as
- reference category.
-
- By playing with the coding scheme you may be able to probe other interesting
- relations. For instance, the somewhat modified scheme shown below would
- allow you to contrast adjacent levels (any j against j-1):
-
- Value for Coding for dummy variables
- NUMP NUMP2 NUMP3 NUMP4 NUMP5 Level
- 0-1 0 0 0 0 1
- 2-3 1 0 0 0 2
- 4-5 1 1 0 0 3
- 6-7 1 1 1 0 4
- >=8 1 1 1 1 5
-
- Now back to your question, the use of dummy regressors as shown above
- would allow you to model non-linear trends in dose-response between your
- predictor and the outcome. It would also allow you to spot sudden changes
- in effect, such as a threshold, as you mention. In fact, with the latter
- scheme you would use the Wald statistics for each coefficient to detect
- a significant difference in two consecutive ORs.
-
- E. Franco, Univ. Quebec (imr2027@iaf.uquebec.ca)
-